基于自然邻域和数据引力的多标签不平衡数据过采样方法

PDF下载HTML

Title:

Multi-label imbalanced data oversampling based on natural neighborhood and data gravity

作者:

刘志强, 谭浩宇, 韩奥坤, 王炜清, 严远亭, 张燕平

安徽大学 计算机科学与技术学院, 安徽 合肥 230601

Author(s):

LIU Zhiqiang, TAN Haoyu, HAN Aokun, WANG Weiqing, YAN Yuanting, ZHANG Yanping

College of Computer Science and Technology, Anhui University, Hefei 230601, China

关键词:

多标签; 不平衡学习; 过采样; 自然邻域; 数据引力; 标签分配; 类重叠; 分类

Keywords:

multi-label; imbalanced learning; oversampling; natural neighborhood; data gravitation; label assignment; class overlap; classification

分类号:

TP311

DOI:

10.11992/tis.202505019

文献标志码:

2026-2-4

摘要:

在处理多标签不平衡数据分类问题中，过采样方法是主流技术之一。然而，如何设计有效的采样策略以捕捉样本局部分布信息，同时避免合成过程引入重叠样本而导致类间区分度降低，始终是过采样面临的关键挑战。针对该挑战，提出了一种基于自然邻域和数据引力的多标签不平衡数据过采样方法。该方法首先基于特征空间构建自然邻域结构，以自适应学习样本的局部分布信息。其次利用标签相似性来引导辅助样本选择，为相对安全的辅助样本赋予更高的权重，降低类重叠风险。最后建立数据引力模型构建动态标签分配机制，自适应生成标签信息，避免固定标签分配规则可能引发的类间冲突问题。在14个不平衡数据集上的实验表明，所提算法相较于SOTA方法在3个主要指标上均取得了更优的性能表现。

Abstract:

In multi-label imbalanced data classification, oversampling has emerged as a mainstream technique. However, how to design effective sampling strategies that capture the local distribution information of samples while avoiding the introduction of overlapping samples during the synthesis process, and reducing the inter-class separability, remains a key challenge for oversampling methods. To this end, we propose a novel multi-label oversampling method based on natural neighborhood and data gravitation. Firstly, the method constructs adaptive natural neighborhood structures in feature space to capture local distribution information. Then, it employs label similarity to guide auxiliary sample selection, assigning higher weights to relatively safe auxiliary samples to mitigate class overlapping risk. Finally, it constructs a dynamic label assignment mechanism with the data gravitation model to generate label information, and avoiding the possible inter-class conflicts inherent in fixed label allocation rules. Experimental resultson 14 imbalanced datasets demonstrate that the proposed algorithm outperforms state-of-the-art methods in three performance metrics.